The short age of dragon riders*
Context
Since approximately August of this year, the AI models available to the public have passed a threshold: in a proper harness, they are stronger than most theoretical physicists (myself included) at most technical tasks. What counts as “technical” exactly is blurry (and likely to expand) but I take it to include the bulk of what junior theoretical physicists were doing day to day during their PhDs.
Concretely, the models are now able to carry out complex analytical calculations, work out lengthy proofs, turn well-posed problems into code, all with minimal input. They iterate, correct their errors, launch simulations to verify their conjectures, and summarize their work. On my own problems, I have already seen the impossible being done several times, starting with Fable 5 in May (I reported some on X and LinkedIn). Those who deny this progress have just not tried hard enough (with the latest models), or are blinded by their ego.
As of November 2026, the best public models still lack research taste. They execute beautifully when given a clear goal but are worse than the best researchers at coming up with good questions, or at having the sparks of creativity needed to get out of a dead end. They do not have as much vision as the best humans.
This situation is unlikely to persist. Even if the whole internet has been exhausted to pre-train models, the availability of human data is no longer the bottleneck. Thanks to reinforcement learning with verifiable rewards (RLVR), models can improve simply by solving new tasks, and are in no way bound by human intelligence.
At least in the immediate future, progress is thus predictable. First, the internal models at OpenAI and Anthropic have now solved countless open problems, showing that they are indeed a generation ahead of what the public currently has access to (which is already superhuman at many tasks). Second, computing power, the only bottleneck, has been increasing steadily, far beyond what was used to train the current best models. Unless the scaling laws surprisingly break down, we have two generations of improvements ahead of us without doing anything new. On top of this, recursive self-improvement is under way, at least in the weak sense that strong models already substantially help design the training environments of their successors. In a nutshell, as long as the datacenters survive kinetic strikes or a grassroots Butlerian Jihad, progress is not going to stop.
Like most scientists, I am conflicted. All the friction is gone; problems melt before we can even understand the solution. It is great, and also scary, leading some in the community to hate the models. I have always been a techno-optimist absolutist. It would be a betrayal to flip now that the wave of progress touches my craft. No, even if it burns me, I still think we should steal the fire!
Granted, the fire, this time, may come with existential risks for humanity. I think these risks are real, and to leading order, better understood by the nerds in Silicon Valley than by skeptics like Yann LeCun. Alignment research is certainly important, and if slowing down training helps reduce existential risk, then we probably should.
But I think it is too easy to reject AI progress for wrong, self-serving reasons, under the guise of a legitimate concern for mankind. And thus, while one should take these risks seriously, it is important to also think about what is left for us in case they are solved.
It is still too difficult for me to envision what the steady state may look like, or whether there will even be one. But I would like to discuss what I think we should do (or at least what I am doing) during the transient period of symbiotic human-AI discovery.
Grief
The first step is to quickly go through the five stages of grief. Mathematicians, as Sylvia Serfaty recently noted in her talk at the IHP, at least seem to have gone through the denial phase. Physicists are likely only three months behind.
Prominent mathematicians, even those who were dismissing AI only three months ago, are no longer denying its power. However, the most visible public reaction, initiated by 25 Fields medalists, is terrible in a way that can only be explained by shock. The best problem solvers alive are now claiming that results, proofs, technical breakthroughs, which in many cases gave them their status and fame, were not really the goal. The goal was the friends we made along the way: the serendipitous discoveries, the mathematical community that got built, the long hours spent pondering in front of a blackboard. This enjoyment, this human understanding, for the community and by the community, is what had value all along.
I disagree. I think we deserve all the bad things that will happen to us if we react like such sore losers. Serendipitous discoveries are important, but first, it is unclear that AIs will necessarily lead to fewer, and second, the end results clearly mattered and still matter. Seeing the process as an end in itself, rather than a means to an end, is profoundly self-serving, and sadly typical of academia.
We have to accept that many things we were doing, sometimes with joy, are already gone. I have spent days massaging an equation to bring it into a form that worked, proud of my virtuosity. I have made complicated calculations with pen and paper, covering entire notebooks to get an answer I could trust. I have spent weeks turning computations into code, and optimized the implementation over months, to be the first to extract certain results from theory. I have fine-tuned plots for hours until they looked pretty. Science was slow; all these steps were important and had to be taught. They were sometimes tedious, sometimes meditative, and tied to our identity. I was good at this, and enjoyed at least a fraction of it. But even if the models do not make any progress, which as I argued is highly unlikely, this is gone. Spending too much time on what is gone will soon be unprofessional, like bringing a coloring book to work.
The temptation of vegan science
Since AI is now needed to push the frontier of research and understanding, those who do not want to use it sometimes argue that the output of research is researchers themselves (and not so much the results they produce). Hugo Duminil-Copin, a French Fields medalist, explains (in a recent talk at the Société Mathématique de France) that no one really cares about or understands his proofs outside of a few experts (which is mostly true, although I personally enjoyed his lectures as a non-expert). As a result, he concludes, his main value for society lies in showing how far the human mind can be pushed with sustained effort and dedication. Hugo Duminil-Copin is an athlete of the mind, an example for the rest of society. If this is true, then indeed Duminil-Copin can never be replaced by AI, and he is right not to use AI too much. He has said he will digest the AI proofs to understand them and pass them on, a bit like chess players studying engine lines, while continuing his own research without AI.
Machines may progress to the point where this view (which I find self-serving) is the only reasonable one. But during the transient, I cannot accept it. At least after they are done with their PhDs, researchers are no longer the product of research; what matters is what they do.
I hold the naive view that the goal of research is to push toward new results, knowledge, applications, understanding that can be shared. Discoveries are usually a good thing, actually. I think there will still be joy in discoveries even if it is not only humans that push them. If the old ways are gone, at least I want to push to the stars!
Riding the dragon
As long as there are problems to solve, and as long as we bring something to their resolution, I think we should ride the AI dragon. We should enhance ourselves, move up in abstraction, and let the dragon deal with what it does better than us.
To do great things, we need to forgo some understanding of what is going on at lower levels, to focus on higher ones. This is still controversial in my field, and difficult to accept, but this has happened before: we no longer code in assembly, compute integrals manually with quadratures, or extract square roots with Newton’s method. The level of understanding we had was adapted to our tools; now that the tools have changed, what we need to understand fully also changes.
I doubt that riding the dragon can be a program for the steady-state ASI scenario (if there is one), but it can be a program for now. The dragon is getting bigger every three months. It may ride itself at some point, but so far few can ride it properly. Those who do not ride it now may never be able to ride it in the future, at least not to find new things.
Riding the dragon is fun but also exhausting, and in some ways humiliating. The best way to learn is just to do it, and to spend many hours looking the beast in the eye. It is not yet clear how to balance the time spent in front of the screen, actively prompting, with the time spent pondering the output during a walk in the park, to catch up with what the dragon found, and be able to productively point it in another direction next. I suspect we will never know what the right way is, as it changes too fast for us to find it. I certainly do not have a silver bullet, and I think some level of intellectual discomfort is worth it. Like many others, I try to push myself to the edge of psychosis, getting the most out of it while remaining sane. Discussing the output with other humans (ideally other dragon riders) also helps stay grounded. Sleeping the usual amount is vital, but increasingly difficult.
As in a bubble, I suspect the value of good dragon riders will increase exponentially until it collapses dramatically. At some point it will not be possible to raise the bar, stay above the wave, and it is plausible that nothing biological will be able to give valuable insights to the silicon. The dragon will know what to do. But we live now, not in three years.
As researchers, all that we hoped to be able to solve and find in a lifetime is probably going to be done in the next few years. We can remain bitter spectators of these great achievements, or ride the dragon to take part in them, for the honor of the human mind.
* I have given a bunch of talks defending this vision in the past months, the most recent one at the IHP (slides, video).